1 / 46

FlexiCheck: an Adaptive Checkpointing Architecture for Energy Harvesting Devices

FlexiCheck: an Adaptive Checkpointing Architecture for Energy Harvesting Devices. Priyanka Singla , Shubhankar S. Singh and Smruti R. Sarangi School of Information Technology Indian Institute of Technology Delhi, India.

mitton
Download Presentation

FlexiCheck: an Adaptive Checkpointing Architecture for Energy Harvesting Devices

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. FlexiCheck: an Adaptive Checkpointing Architecture for Energy Harvesting Devices Priyanka Singla, Shubhankar S. Singh and Smruti R. Sarangi School of Information Technology Indian Institute of Technology Delhi, India This work has been partially supported by the Semiconductor Research Corporation (SRC)

  2. Outline • Introduction • Existing Work • Mathematical Modelling • Proposed Model – FlexiCheck • Evaluation • Conclusion Introduction

  3. Difficult to replace • Large form factor Use ambient energy, which otherwise goes wasted

  4. Energy Harvesting Devices (EHDs) Convert ambient energy into electrical energy and store in a small storage unit (capacitor) ISSUES!! small Solar Energy • Uncertainambient energy • Rate of charging is much less than the rate of discharging Different sources have different characteristics 1. Time Vibrational Energy Vibrational Power Solar 1. Image Source: http://www.schoolphysics.co.uk

  5. Normal Execution: Battery based device. • EHD Execution: Capacitor based device. Normal ExecutionvsEHD Execution

  6. Normal Execution

  7. Program Finishes Successfully Program Begins Normal Execution EHD Execution

  8. Program Finishes Successfully Program Begins Normal Execution EHD Execution Energy Store Program Failure Program Begins Storage Drained Out

  9. Program Finishes Successfully Program Begins Normal Execution The cycle continues EHD Execution Energy Store Program Failure Program Begins Storage Drained Out

  10. Program Finishes Successfully Program Begins Normal Execution EHD Execution Energy Store Checkpoint Program Begins

  11. 11 Wait and Recharge Energy Store Program Finishes Successfully Restore

  12. Device’s Execution Cycle

  13. Outline • Introduction • Existing Work • Mathematical Modelling • Proposed Model – FlexiCheck • Evaluation • Conclusion Existing Work

  14. Checkpointing Techniques • STATIC: Programmer or compiler intervention. • Conservative: Oblivious to ambient energy • E.g., Dino[1], Ratchet[2], Chain[3] • FLEXICHECK: ADAPTIVE (Dynamic + Learning) • No inputsfrom programmer • Architecture predicts the profile, dynamically adapts the checkpointing strategy • DYNAMIC: Use energy monitoring by hardware • Runtimedecisions: Checkpoint when energy is about to exhaust • E.g., Hibernus[4], Mementos[5] • LEARNING: Learn ambient profile offline, ambient energy not monitored • at runtime • E.g., MDP[6] 1. B. Lucia and B. Ransford, “A simpler, safer programming and execution model for intermittent systems,” PLDI 2015. 2. J. Van et al., “Intermittent computation without hardware support or programmer intervention.” in OSDI, 2016. 3. A. Colin and B. Lucia, “Chain: tasks and channels for reliable intermittent programs,” OOPSLA 2016. 4. D. Balsamo et al., “Hibernus: Sustaining computation during intermittent supply for energy-harvesting systems,” IEEE ESL, 2015. 5. B. Ransford et al. , “Mementos: System support for long-running computation on rfid-scale devices,” ASPLOS 2011 6. Z. Ghodsi et al., “Optimal checkpointing for secure intermittently-powered iot devices,” in ICCAD, 2017.

  15. Outline • Introduction • Existing Work • Mathematical Modelling • Proposed Model – FlexiCheck • Evaluation • Conclusion Mathematical Modelling

  16. Mathematical Model of the Problem • Model • Optimization Formulation • Modified Formulation: • Quadratically Constrained Linear Program (QCLP)

  17. Ambient Energy Solar Energy Power (mW) Vibrational Energy Time (ms)

  18. Ambient Energy Program • Subset Creation Create Disjoint Subsets S2 S1 Sm should not be more than the device can provide Feasible Points Checkpoints can be taken only at boundaries, not within a subset

  19. Ambient Energy • Subset Creation Sm S2 S1 • Checkpointing Cost • Recharging Time Wi Xi=1 • Entire Volatile memory (SRAM and registers) • Energy (Ec) and Time (Tc) Wait (Wi) is always preceded by a checkpoint (Xi) Wait and recharge anytime, not only at end.

  20. Mathematical Model of the Problem • Model • Optimization Formulation • Modified Formulation: • Quadratically Constrained Linear Program (QCLP)

  21. Optimization Formulation Checkpointing Subset exec Subset exec Xi=0 Xi=1 Xi=1 Xi=1 22

  22. Optimization Formulation Checkpointing Checkpointing Waiting Subset exec Subset exec Wait Time Xi=0 Xi=1 Xi=1 Xi=1 23

  23. Optimization Formulation Variables Subject to: = subset execution energy • Both objective function and constraints highly non linear • Computationally intractable • Reduce complexity Energy spent in checkpointing Initial stored energy Energy gained while waiting 24

  24. Mathematical Model of the Problem • Model • Optimization Formulation • Modified Formulation: • Quadratically Constrained Linear Program (QCLP)

  25. Approximations to reduce the complexity Wait time and execution time can be computed in the same way: only keeping the track of the interval S2 S1 Sm Rate of discharge is much more than rate of charging. W Fixed number of instructions, similar time and energy for all the intervals, choose max No need to keep a track of per interval energy or time

  26. = 0 Modified Optimization Formulation Linear Program = 1 Subject to: 1 𝒩 𝒩-1 2 3 4 𝒩-2 6 5 Modified Optimization Formulation: QCLP : Quadratic Constraints 27

  27. Outline • Introduction • Existing Work • Mathematical Modelling • Proposed Model – FlexiCheck • Evaluation • Conclusion Introduction

  28. FlexiCheck • Optimal solution for a particular ambient profile. • Does not deal with variations in the profile Why FlexiCheck? • Implement an energy predictor in hardware • Learn some parameters from QCLP • Based on the prediction, take actions in software Idea and its relation to QCLP

  29. Ambient Energy Source Vibrational source, vary sinusoidally at ms granularity S (Power mW) Define a threshold St (QCLP), S >St good to recharge Time (ms) St

  30. Decision Logic Decision Logic Ambient Energy (S) Action • Action A is a tuple (Ex, Ck) • Ex=1: Execute, else wait • Ck=1: Checkpoint (St, Bt) St => Bmax S (Power mW) Using QCLP Bt Bmin Time (ms)

  31. Implementation Characterize Susing periods, denoted as (h,d) pairs • Space efficiency: • Few bits for (h,d). • Need not store values for all the points in a period. S (Power mW) • Energy efficiency: • No need to compare S with St, in every clock cycle, only if new (h,d). • Since periodic, so (h,d) will be maintained, and hence the frequency of computing actions decrease. Time (ms)

  32. Implementing in Hardware Train Decision Logic Find (h,d) of the period Prediction

  33. Read (h,d) Implementing in Hardware Train Prediction Find action

  34. Learning Thresholds (St, Bt) Computed using QCLP St 𝒩 𝒩-1 From QCLP: For each interval, , we have a tuple ((, ), ) Bt Apply linear regression on all the points, get the intercepts

  35. Outline • Introduction • Existing Work • Mathematical Modelling • Proposed Model - FlexiCheck • Evaluation • Conclusion Evaluation

  36. Designed in Verilog, used Cadence Genus Tool, 130nm Technology Synthesis Results:Area, Time and Energy

  37. AR RSA Benchmarks CEM LED CF

  38. Platform Experimental Setup Processor System Parameters 39

  39. ExperimentsPerformed • Time and energy comparison with variable energy • Number of checkpoints

  40. Performance FlexiCheck is within 3-8% of QCLP Chain: 6-50 X slower Dino: 6-36X slower Mementos: 2-5X slower • A. Colin and B. Lucia, “Chain: tasks and channels for reliable intermittent programs,” OOPSLA 2016. • B. Lucia and B. Ransford, “Dino: A simpler, safer programming and execution model for intermittent systems,” PLDI 2015. • B. Ransford et al. , “Mementos: System support for long-running computation on rfid-scale devices,” ASPLOS 2011.

  41. Energy FlexiCheck ~20% less energy than QCLP Chain: 10-30X more Dino: 6-25X more Mementos: 2-5X more • A. Colin and B. Lucia, “Chain: tasks and channels for reliable intermittent programs,” OOPSLA 2016. • B. Lucia and B. Ransford, “Dino: A simpler, safer programming and execution model for intermittent systems,” PLDI 2015. • B. Ransford et al. , “Mementos: System support for long-running computation on rfid-scale devices,” ASPLOS 2011.

  42. # Checkpoints FlexiCheck has least number of checkpoints • A. Colin and B. Lucia, “Chain: tasks and channels for reliable intermittent programs,” OOPSLA 2016. • B. Lucia and B. Ransford, “Dino: A simpler, safer programming and execution model for intermittent systems,” PLDI 2015. • B. Ransford et al. , “Mementos: System support for long-running computation on rfid-scale devices,” ASPLOS 2011.

  43. Outline • Introduction • Existing Work • Mathematical Modelling • Proposed Model - FlexiCheck • Evaluation • Conclusion Conclusion

  44. Conclusion

  45. 46

  46. FlexiCheck: an Adaptive Checkpointing Architecture for Energy Harvesting Devices Priyanka Singla, Shubhankar S. Singh and Smruti R. Sarangi School of Information Technology Indian Institute of Technology Delhi, India This work has been partially supported by the Semiconductor Research Corporation (SRC)

More Related